Solution · Seal
Frontier-class AI with no route to the internet.
Air-gapped AI is an LLM deployment on a network physically or logically isolated from the internet: inference, retrieval, and fine-tuning run entirely inside the enclave, with no telemetry, no license-check callbacks, and no external party in the chain. It is the deployment pattern for CMMC and ITAR programsITAR flow-downs, Controlled Goods environments, and any organization whose secrets are worth a sealed network.
Who it’s for
When the network boundary is the compliance boundary.
Defense contractors handling CUI under CMMCCMMC flow-downs, firms with ITAR-controlled technical data, manufacturers protecting process trade secrets, and institutions running classified-adjacent programs all share one requirement: certain data must never touch a system outside the authorization boundary. Cloud AI fails that requirement by definition. Air-gapped open-weight deployment passes it by construction — the model is a static file inside the enclave, and the audit answer to “does data leave?” is no.
The subtle risks live in the stack, not the model: serving frameworks that report usage telemetry, GPU drivers that check licenses, container registries that assume connectivity. Our build process vets, pins, and mirrors every component so the environment runs — and updates — without ever needing a route out.
The engagement
How an air-gapped build works.
- Boundary assessment. We map your controlled-data classes, authorization boundaries, and the workloads that justify the enclave.
- Sealed-stack engineering. Every component — serving framework, drivers, registries, monitoring — vetted, pinned, and mirrored inside the boundary.
- Controlled ingest. Weights and updates enter by checksummed, change-managed media transfer; nothing enters or leaves at runtime.
- Enclave operations. Per-program access controls, audit logging, and runbooks your cleared personnel operate — with our support structured to match your security requirements.
Deliverables: a sealed, documented serving environment; pinned and checksummed model weights; an offline update procedure integrated with your change management; and evaluation results for every model admitted to the enclave.
Questions we get
Frequently asked questions
Can a modern LLM really run with no internet connection at all?
Yes. Open weights are static files. Once downloaded, checksummed, and transferred by controlled media, the full stack — inference, retrieval over internal documents, and fine-tuning — runs on a sealed network indefinitely. Nothing in a properly built serving stack requires a route out; the common failure mode is unvetted components that phone home for telemetry or license checks, which is exactly what our hardening pass removes.
How do model updates reach an air-gapped environment?
By the same discipline as any controlled software: new weights are downloaded and checksummed outside the enclave, scanned, approved through your change-management process, and moved in by controlled media transfer. Because upgrades happen on your schedule, the version in production is always one your own evaluation harness approved.
What do we lose by air-gapping compared to a connected deployment?
Live web retrieval — and that is essentially all. Reasoning, drafting, code work, and retrieval over your internal corpus are unaffected, and long-context models (up to 1M tokens) carry entire document sets in a single pass. For CUI, ITAR-controlled data, and trade secrets, the exchange is decisively favourable: the network boundary becomes the compliance boundary.
Seal the boundary. Keep the capability.
The two-week sovereignty assessment maps your controlled-data classes and workloads, and hands you a written enclave architecture with a real cost model.
Book a sovereignty assessment